To address the limitations of existing GPT-based password guessing models, which suffer from insufficient contextual awareness and generation quality bottlenecks when modeling complex hybrid structures, as well as the semantic chunk loss issue in pattern-guided models caused by reliance on traditional PCFG field segmentation, this paper proposes ChunkPassGPT, a chunk-level pattern-guided generative password guessing model. Inspired by the Byte Pair Encoding (BPE) algorithm, ChunkPassGPT introduces a novel password segmentation method that divides passwords into high-frequency, semantically coherent chunk units. These chunks are then used to construct chunk-level pattern sequences, which serve as guiding information fed into the GPT architecture. Through a concatenated training paradigm involving both “pattern sequence – password sequence” pairs, the model simultaneously learns structural patterns and content generation rules, significantly enhancing its ability to model cross-type semantic units. Furthermore, during the generation process, the model incorporates the frequency distribution of different chunk-level patterns in the training set, employing a pattern frequency-weighted probabilistic generation strategy to enhance the alignment between the generated results and real user behavior. Experimental results show that ChunkPassGPT improves the hit rate by 23.65% and reduces the duplication rate by 13.07%.